The AI molecule hunt that targets machinery a virus borrows

An antiviral search can begin with a protein that belongs to us. Researchers at Korea University are using AI to look for compounds that inhibit deoxyhypusine synthase, or DHPS, a human enzyme involved in machinery that viruses can exploit.[2]

Their study was published on 5 August and publicised on 19 October. The result is a set of early enzyme-inhibiting fragments, with two showing measurable half-maximal inhibition in laboratory tests. It has not established an antiviral treatment.[1][2]

Why look on the host side?

Viruses depend on cells to reproduce. An independent 2025 study used genome-wide CRISPR screens across nine respiratory viruses to investigate the human genes supporting infection. Its shared dependency network included the machinery for modifying the protein eIF5A.[3]

That gives researchers a different search direction: a cellular process shared by several infections might offer a broader target than a feature unique to one virus. It also raises a demanding question. Can an intervention interfere with what the virus needs while preserving enough of the host’s normal work?

DHPS participates in activating eIF5A, a factor involved in making proteins. Activation requires a two-stage chemical modification called hypusination: DHPS performs the first stage and a second enzyme, DOHH, completes it. The target is therefore part of an existing cellular process, not a switch installed by a virus.[4]

From a predicted fit to an enzyme test

The Korean team compared 13 methods for predicting binding scores and ten approaches to generating molecules. It combined several kinds of prediction model, then used that scoring system to guide a reinforcement-learning generator. The intended benefit is a more productive shortlist for physical testing.[2]

The underlying training labels were calculated binding-energy estimates for 1,920 fragments, not measured antiviral effects. For laboratory work, the researchers obtained six commercially available compounds selected from the proposals: three predicted active and three predicted inactive.[1]

Tests covered concentrations from 1 to 100 micromolar. Two compounds had IC50 values of 31.2 and 49.1 micromolar. A third produced only 10.9% inhibition at 100 micromolar; the three predicted inactive compounds showed none across the tested range.[1]

IC50 is the concentration needed to reduce the measured activity by half under that assay’s conditions. It is a useful description of an experimental starting point, not a patient dose or a measure of how well the compound treats an infection.

The team describes the hits as small fragments with chemical frameworks different from the familiar inhibitor GC7. That leaves space for subsequent chemical modification, with the aim of improving binding and other properties together.[2]

The assay is one link in a longer chain

Independent work on measuring hypusination explains why the next steps are substantial. A cell-free assay can show that a compound changes an enzyme reaction. It cannot establish whether the compound enters intact cells, reaches the right location or remains stable long enough to work.[4]

That research also developed a way to measure the modified eIF5A protein itself, using an antibody-based readout. It illustrates how researchers can ask a more direct downstream question after finding an apparent enzyme inhibitor: did the expected protein modification actually change?[4]

The Korean study leaves cellular activity, detailed inhibition mechanisms, selectivity and toxicity for further work. Its contribution is a chemically testable starting point for that programme.[1]

The attraction of the host-directed approach is a possible shared vulnerability across viruses. The difficulty follows from the same fact: the machinery also has a job in a healthy cell. AI can help search the chemical possibilities; the experiments have to establish which possibilities are useful.

Report an error or suggest a correction